14th International Conference on Information Fusion

Chicago, Illinois, USA, July 5-8, 2011

Change Detection from Remote Sensing Images

Based on Evidential Reasoning
Zhun-ga Liu1,3 , Jean Dezert2 , Gregoire Mercier3 , Quan Pan1 ,Yong-mei Cheng1
1. School of Automation, Northwestern Polytechnical University, Xian, China, Email: liuzhunga@gmail.com
2. Onera - The French Aerospace Lab, F-91761 Palaiseau, France, Email: jean.dezert@onera.fr
3. Telecom Bretagne, Technopole Brest-Iroise, 29238, France, Email: Gregoire.Mercier@telecom-bretagne.eu

AbstractTheories of evidence have already been applied more

or less successfully in the fusion of remote sensing images. In the
classical evidential reasoning, all the sources of evidence and
their fusion results are related with the same invariable (static)
frame of discernment. Nevertheless, there are possible change
occurrences through multi-temporal remote sensing images, and
these changes need to be detected efficiently in some applications. The invariable frame of classical evidential reasoning
cant efficiently represent nor detect the changes occurrences
from heterogenous remote sensing images. To overcome this
limitation, Dynamical Evidential Reasoning (DER) is proposed
for the sequential fusion of multi-temporal images. A new state
transition frame is defined in DER and the change occurrences
can be precisely represented by introducing a state transition
operator. The belief functions used in DER are defined similarly
to those defined in the Dempster-Shafer Theory (DST). Two
kinds of dynamical combination rules working in free model
and constrained model are proposed in this new framework for
dealing with the different cases. In the final, an experiment using
three pieces of real satellite images acquired before and after
an earthquake are provided to show the interest of the new
approach.

Keywords: Evidence Theory, Change Detection, Dynamical

Evidential Reasoning, DST, DSmT, Remote Sensing.
I. I NTRODUCTION
Information fusion resulting from multi-temporal and multisources remote sensing images remains an open and important
problem [1]. The remote sensing images can be quite different
in their modality [2]: orbits may be ascending and descending,
parameters of acquisitions may differ from one image to
another even when the two acquisitions are issued from the
same sensor. That is why, for change detection purpose, the
use of the difference image is not an appropriate point of view
due to the number of false alarms it induces. The situation is
even worse on high resolution images over urban areas since
many building appears differently in the two images to be
compared due to the geometry of sensors, the perspective, the
light condition and shadows... Hence the comparison of the
classified images seems to be more appropriated. But, this
yields to deal with uncertain, imprecise and even conflicting information. Evidence theories including Dempster-Shafer
Theory (DST) [3] and Dezert-Smarandache Theory (DSmT)
[4] are good for dealing with such information, and they
have been applied for remote sensing applications [1, 5, 6].
In past works, a particular attention was paid to obtain very
specific results for decision-making support through efficient

978-0-9824438-3-5 2011 ISIF

fusion of sources of evidence. Thus many works focused

mainly on the redistribution of the conflicting beliefs [7, 8].
These combination approaches can be called static approaches,
since they work under the assumption that the frame, on
which is based the decision-making support, is temporally
invariable in the fusion process. However, in the fusion of
the multi-temporal remote sensing images, unexpected change
occurrences can arise in some parts of the images. So both
the image classification and change detection will be involved
together. The classical combination rules in evidence theories
provide specific classification results in the invariable parts of
images, but it cannot precisely detect the change occurrences
in the variable parts.
Therefore, a dynamical evidential reasoning (DER) working
under the condition that the frame does not necessarily remain
invariable in the fusion is proposed. In this paper, a new frame
called State-transition power-set is defined, and the change
occurrences among different hypotheses can be precisely represented by the state transition operator in this frame. The
dynamical belief functions Bel(.), plausibility functions P l(.)
and pignistic probability BetP (.) [9, 10] are defined similarly
as in DST. Dynamical combination rules are then proposed to
work either in the free model, or in the constrained model. The
free model is well adapted when no prior knowledge is known
on elements of the frame. The constrained model can be used
if some integrity constraints about the change occurrences are
known. The dynamical approach improves the performance
of the classification of areas and estimation of the changes
through the fusion of multi-temporal sources of evidence. Our
proposed approach is finally applied for the fusion of three
sequential pieces of satellite images acquired before and after
an earthquake.
II. DYNAMICAL EVIDENTIAL REASONING APPROACH
A. A brief review of DSmT
We need to introduce briefly DSmT framework because
the Dynamical Evidential Reasoning approach shares some
common ideas with DSmT, in particular the necessity to deal
with hybrid models of the frames in some applications for
changes detection. The purpose of DSmT is to overcome the
limitations of DST [3] mainly by proposing new underlying
models for the frames of discernment in order to fit better
with the nature of real problems, and proposing new efficient
combination and conditioning rules. In DSmT framework, the

977

elements i , i = 1, 2, . . . , n of a given frame are not

necessarily exclusive, and there is no restriction on i but their
exhaustivity. The hyper-power set D in DSmT is defined as
the set of all composite propositions built from elements of
with operators and . For instance, if = {1 , 2 },
then D = {, 1 , 2 , 1 2 , 1 2 }. A (generalized) basic
belief assignment (bba for short) is defined as the mapping
m : D [0, 1]. The generalized belief and plausibility
functions are defined in almost the same manner as in DST.
Two models1 (the free model and hybrid model) in DSmT
can be used to define the bbas to combine. In the free DSm
model, the sources of evidence are combined without taking
into account integrity constraints. When the free DSm model
does not hold because the true nature of the fusion problem
under consideration, we can take into account some known
integrity constraints and define bbas to combine using the
proper hybrid DSm model. All details of DSmT with many
examples can be easily found in [4].
B. The space of state transitions in DER
DSmT has been already applied for the fusion of multitemporal satellite images in [1]. However, the classical frame
used is not well adapted for measuring the changes among its
elements. The conjunctive elements (intersections) in hyperpower set D represent either the overlap between hypotheses
in free DSm model, or the conflict produced by the conjunctive
combination in Hybrid DSm model when this is an integrity
constraint. Actually, the conjunction A B is unable to
characterize the transition A changing to B (denoted A B),
or B changing to A (denoted B A). Therefore, if we
need to distinguish two possible state transitions for changes
detection, we need to define new operator and cannot use
the classical conjunctive/intersection operator as in classical
approaches. We propose the state transition operator changing
to, denoted , satisfying the following reasonable conditions:
(C1) Impossible (forward) state-transition

obviously. tA,A = A A represents a particular invariable

transition. A chain of transitions 1 2 n will be
denoted t1,2,...,n . A transition i (j k ) will be denoted
ti,jk , etc.
In the theories of belief functions (DST, DSmT or Transferable Belief Model [9]), the result of the fusion of sources
of evidence defined on a same frame of discernment , is
obtained by a given rule of combination relatively to a fusion
space G , where G can be either the classical power set
2 , a hyperpower-set D , or a superpower set (the power set
of the refined frame) depending on the theory used. In this
paper, we propose to use another fusion space (the space of
transitions), denoted T , in order to deal explicitly with all
possible state transitions we want to detect.
Firstly, the transition frame is given by:
1n = 1 2 . . . n
= {tX1 ,X2 ,...,Xn |Xi i , i = 1, 2, . . . , n}
where i is the frame associated with the i-th source and
is the Cartesian product operator.
In this paper, we assume to work in a more simple case
where all frames i , i = 1, 2, . . . , n are the same and equal
to , and where G = 2 . In other words, we will work with
the simpler space denoted Tn and defined by
n times

Tn = 21n

Tn can be called state transition power-set, which is composed

by all the elements in 1n with the union operator . We
define as componentwise operator in the following way:
tX , tY Tn ,

A,
(C3) Distributivity of w.r.t.
(A B) C = (A C) (B C)
(C4) Distributivity of w.r.t.
A (B C) = (A B) (A C)
(C5) Associativity of state-transition
(A B) C = A (B C) = A B C.
For notation convenience, a (state) transition A B will be
denoted tA,B . It is important to note that the order of indexes
does matter because tA,B 6= tB,A in general, but if A = B
1 Actually, Shafers model, considering all elements of the frame as truly
exclusive, can be viewed as a special case of hybrid model.

tX tY = tXY .

(1)

Following conditions C1-C5, we note that generally

t(X1 ,X2 ,...,Xn ) t(Y1 ,Y2 ,...,Yn ) = t(X1 ,X2 ,...,Xn )(Y1 ,Y2 ,...,Yn )
6= tX1 Y1 ,X2 Y2 ,...,Xn Yn

A,
(C2) Impossible (backward) state-transition

z
}|
{
= 2 . . .

If Xi 6= Yi ; Xi , Yi are singletons, t(X1 ,X2 ,...,Xn )(Y1 ,Y2 ,...,Yn )

indicates only two kinds of possible transitions:
tX=(X1 ,X2,...,Xn ) or tY =(Y1 ,Y 2,...,Yn ) , whereas the element
tX1 Y1 ,X2 Y 2,...,Xn Yn represents 2 2 2 = 2n
kinds of possible transitions. It is obvious that they are quite
different, and tX1 Y1 ,X2 Y2 ,...,Xn Yn is much more imprecise
than t(X1 ,X2 ,...,Xn )(Y1 ,Y2 ,...,Yn ) .
As we see, the important and major difference between the
classical approaches (DST, TBM, DSmT) and DER approach
is the choice of the fusion space we are working with. With
DST, TBM or DSmT, the fusion space we work with is
always the same (independent of the number of sources) as
soon as the sources are defined with respect to same frame
, whereas with DER approach the fusion space is always
increasing with the number of sources, even if the sources are
all referring to the same frame . This of course increases
the complexity of DER approach, but this is the price to
pay to identify and estimate the possible change occurrences

978

in remote sensing images sequence as it will be shown in last

section of this paper. Clearly, DST and DSmT are not well
adapted for detecting change occurrences. For example, the
transition tA,B,A from state A in source 1, to state B in source
2, and back to state A in source 3 cannot be represented in
the fusion space proposed with TBM, DST, nor in DSmT.

focal element. For example, m(t(1 2 ),3 ) = m(t1 ,3 t2 ,3 )

because of the condition (C3). The belief, plausibility functions and the pignistic transformation are defined in DER
similarly as in DST; that is:
X
Bel(A) =
m(B)
(2)
A,BTn ;BA

Example 1: Lets consider = {1 , 2 } with Shafers model,

then 2 = {, 1 , 2 , 1 2 }. We first consider at time 1 the
initial set of invariable transitions defined as follows:
11 = ;
T1 2 ={t , t1 1 , t2 2 , t12 1 2 }
If we want to consider all the possible transitions from time
stamp 1 to stamp 2, one starts with the cross product frame

={, t1,1 , t1,2 , t2,1 , t2,2 ,

t1,1 t1,2 = t1,12 , t1,1 t2,1 = t12,1 ,
t1,1 t2,2 = t(1,1)(2,2) , t1,2 t2,1 = t(1,2)(2,1) ,
t1,2 t2,2 = t12,2 , t2,1 t2,2 = t2,12 ,
t1,1 t1,2 t2,1 = t(1,1)(1,2)(2,1) ,
t1,1 t1,2 t2,2 = t(1,1)(1,2)(2,2) ,
t1,1 t2,1 t2,2 = t(1,1)(2,1)(2,2) ,
t1,2 t2,1 t2,2 = t(1,2)(2,1)(2,2) ,
t1,1 t1,2 t2,1 t2,2 = t12,12 }.
C. Basic definitions in DER
Belief function Bel(.), plausibility function P l(.) and pignistic probability BetP (.) [9, 10] 2 are basic and important
functions for the decision making. They can be also used in
DER approach as well. Indeed, all the elements in 1n composed by the state transitions through the singleton elements
have specific and unique meaning, and they are considered
the singleton elements. All the focal elements in Tn can be
decomposed in the disjunctive canonical form using these singleton elements with the operator , and we call that canonical
2 DSmP (.) transformation proposed in [4] which provides of better
probabilistic informational content than BetP (.) can also be chosen instead.
But DSmP (.) is more complicated to implement than BetP (.) and it has
not been tested in our application for now.

(3)

The interval [Bel(A), P l(A)] is then interpreted as the lower

and upper bounds of imprecise probability for decisionmaking support [3] and the pignistic probability BetP (A)
commonly used to approximate the unknown probability P (A)
in [Bel(A), P l(A)] is calculated by:
BetP (A) =

T2 =212

m(B)

A,BTn ;AB6=

12 = = {t1,1 , t1,2 , t2,1 , t2,2 }

which has || || = 2 2 = 4 distinct elements, and we
build its power set T2 = 212 including its 16 elements as
in the classical way. The element can be interpreted as the
following set of impossible state transitions corresponding to
t, , t,1 , t,2 , t,12 , t1, , t2, and t12, . We recall that the
imprecise elements are derived from application of conditions
C3 and C4 and not from the componentwise union of n-uples
involved in transition indexes. The cardinality of Tn increases
n
with the value of n as |Tn | = 2|| .
The power set of transitions we want to work with for such
very simple example will be given by:

P l(A) =

X
A,BTn ,AB

|A B|
m(B)
|B|

(4)

where |X| is the cardinal of the element X. In DER, the

cardinal of A Tn is the number of the singleton elements
it contains in its canonical form. For example, |t(1 2 ),3 | =
|t1 ,3 t2 ,3 | = 2.
D. Combination rules in DER
The classical combination rules usually work under the
assumption that all the sources of information refer to the
same common frame. The results of existing combination
rules do not deal with unexpected changes in the frame
and they manage the conflicting beliefs without taking into
account these possible changes in the frame. Nevertheless, the
conflicting information is more important than the information
from the agreement for changes detections.
The fundamental difference between the classical approach
and this new Dynamical Evidence Reasoning (DER) approach
is that the frame of the fusion process is considered possibly
variable over time, and the fusion process is adapted for the
changes detection and identification. When the elements of
the frame are invariable, they can be considered as a special
case of changes corresponding to invariable transition. The
dynamical combination rules will be defined by using the state
transition operator to take benefit of the useful information
included in the conflict between sources. The combination
rules work in free model and constrained model as well.
1) Combination rule in the free model of transitions: In the
free model, there is no prior knowledge on state transitions,
and all kinds of changes among the elements are considered
possible to happen. We start with the combination of the two
temporal sources of evidences at first. Let m1 and m2 be
two bbas provided by two temporal sources of evidence over
the frame of discernment satisfying Shafers model. Its
corresponding power-set is 2 = {, 1 , 2 , 1 2 , . . . , }.
m1 (A), A 2 is the mass of belief committed to the
hypothesis A by the source no. 1, and m2 (B), B 2 is
the mass committed to B by the source no. 2. The sources
no. 1 and no. 2 are considered as independent.

979

In this work, we propose to compute the mass of belief of a

forward transition tAk ,Bl = Ak Bl , l k as m(tAk ,Bl ) =
mk (A)ml (B) where k and l are temporal stamps/indexes.
Note that because mk (A)ml (B) = ml (B)mk (A), this mass
of belief also can be associated to the backward transition
tBl Ak = Bl Ak . We assume always working with ordered
products corresponding to forward temporal transitions. By
convention, the bba mi provided by the source no. i is assumed
to be available at time i and before the bba mj provided
by the source no. j. Indexes of sources correspond actually
to temporal stamps. Following this very simple principle, the
conjunctive combination rule in the free model of transitions,
denoted DERf, is defined by:
tX1 ,X2 T2 ,

m12 (tX1 ,X2 ) = m1 (X1 )m2 (X2 )

(5)

where X1 21 and X2 22 .
For simplicity and in our application, we consider that the
frames are all the same; that is 1 = 2 = . . . = n =
. This simple conjunctive rule can be extended easily for
combining n sequential sources of evidence as follows:

Direct joint fusion of n sources: tX1 ,X2 ,...,Xn Tn

2) Combination rule in the constrained model of transitions: In the constrained model of transitions, one knows that
some kinds of changes among the different elements cant
occur according to our prior knowledge. The set , {M , }
can be defined in introducing some integrity constraints as
done in the hybrid model of DSmT. M includes all the
transitions in Ti , i = 1, 2, . . . , n, which have been forced
to be empty because of the chosen integrity constraints in
the model M, and is classical empty set. If the sources of
evidence share the same reliability in the combination, the
conflict among the evidences will be regarded as possible
changes or as empty sets depending on the constraints we
have. The mass of empty sets arising from integrity constraints
can be distributed to the other focal elements. The notation
M
tA,B = t, means that the transition tA,B is equivalent to the
transition t in the underlying model M given the integrity
constraints.

m1n (tX1 ,X2 ,...,Xn ) = m1 (X1 )m2 (X2 ) mn (Xn )

(6)
where Xi 2i , i = 1, . . . , n.
In this free model, the result of the combination is very specific
since all kinds of change occurrences are distinguished in the
results, but the computation burden is very large because of
the large increase of the cardinality of Tn with n.
The following example will show the difference between
DER and DSmT in free model.
Example 2: Lets consider three bbas on = {1 , 2 } as
m1
m2
m3

1
0.6
0
0

2
0
1
0.5

DERDS rule of combination: The mass of empty sets

is proportionally distributed to the other focal elements
similarly to Dempster-Shafers rule here and we denote
this rule by DERDS for short. Of course, the conflicting
mass can also be redistributed by some other ways. The
combination rule DERDS is mathematically defined as
follows:

, Xn 2n 2 , tn Tn , n 2
tn1 Tn1
P
m1n1 (tn1 )mn (Xn )
M

m1n (tn ) =

ttn1 ,Xn = tn

1K

(7)

where K represents the mass of belief committed to the

empty sets (i.e. the degree of conflict) which is given by
X
K=
m1n1 (tn1 )mn (Xn ).
(8)

1 2
0.4
0
0.5

ttn1 ,Xn

With DSm free model:

1) DSmC: m(1 2 ) = 0.6, m(2 ) = 0.4
2) DERf: m(t1,2,2 ) = 0.3, m(t1,2,12 ) = 0.3
m(t12,2,12 ) = 0.2, m(t12,2,2 ) = 0.2
For the singleton elements, based on DERf, one gets:
Bel(.) BetP (.) P l(.)
t1,2,2
0.3
0.60
1
0
0.20
0.5
t1,2,1
0
0.05
0.2
t2,2,1
t2,2,2
0
0.15
0.4
The results of DERf precisely represents the belief of
change occurrences, whereas the elements in DSmC
cannot reflect the process of state transition because of its
invariable frame. In the decision making, it indicates that
the hypothesis t1,2,2 is most possible to happen according
to Bel(.), Pl(.) or BetP(.), which means 1 in source 1
changes to 2 in source 2 and to 2 in source 3. It implies
DSmT is not adapted for the change detection because of
its invariable frame.

When considering the direct combination of n sequential

sources altogether, one has tn Tn , and Xi 2 , i =
1, 2, . . . , n
P
m1 (X1 ) mn (Xn )
M

m1n (tn ) =

tX1 ,X2 ,...,Xn = tn

1K

(9)

where
K=

m1 (X1 ) mn (Xn ).

(10)

tX1 ,X2 ,...,Xn

Remark: The summation introduced in (7) and (9) allows

to take into account the integrity constraints of the model of
the space of transitions as shown in the next example. If all
kinds of transitions among different elements are constrained
to be empty sets, the frame will become Shafers model, and

DERDS will reduce to Dempsters rule.
Example 3: Lets consider the frame = {1 , 2 } with
following two integrity constraints representing the impossible
state transitions M , {t1,2 , t2,2 }. Therefore, due to these

980

integrity constraints, the fusion space T2 , as given in details

in the Example 1, reduces to the simple following set

The three sequential images are clustered in C = 5 groups,

and the classifications are defined as (colors refer to Fig. 1):

T2 = {, t1,1 , t2,1 , t12,1 }

1 ={Dark area: green plant(w1 )or shadow(w2 )}

2 ={Maroon area: bared soil(w3 )}
3 ={Dark gray area: gray building(w4 )}

We also consider the following bbas inputs:

1
2

m1 0.4 0.6
0
m2 0.5 0.2 0.3
According to the underlying hybrid model M, only the
products m1 (1 )m2 (2 ) and m1 (2 )m2 (2 ) take part in the
conflict as: K = m1 (1 )m2 (2 ) + m1 (2 )m2 (2 ) = 0.2.
The conjunctive mass of belief of possible transitions are
given by
m (t1,1 ) = m1 (1 )m2 (1 ) = 0.20
m (t1,12 ) = m1 (1 )m2 (1 2 ) = 0.12
m (t2,1 ) = m1 (2 )m2 (1 ) = 0.30
m (t2,12 ) = m1 (2 )m2 (1 2 ) = 0.18
M

Due to the integrity constraints t1,2 = , t2,2 = , one has

(
M
t1,12 = t1,1 t1,2 = t1,1 = t1,1
M

t2,12 = t2,1 t2,2 = t2,1 = t2,1

Therefore, the mass m (t1,12 ) must be transferred to t1,1 ,
whereas m (t2,12 ) must be transferred to t2,1 only. Finally,
the result given by DERDS rule is
m (t1,1 ) + m (t1,12 t1,1 )
= 0.40
1K
m (t2,1 ) + m (t2,12 t2,1 )
= 0.60
m(t2,1 ) =
1K

m(t1,1 ) =

It shows that the combination rule in constrained model can

provide more specific results than in the free model when the
associated constrained information is known.
III. A PPLICATION ON REAL REMOTE SENSING IMAGES
Three pieces of multi-temporal satellite images are analyzed
by using DER in this experiment. Three QuickBird images
have been acquired before and after the may 21st 2003
earthquake in the region of Boumerdes, Algeria. They have
been corrected geometrically by using SRTM elevation data
and resample by a P+Xs pan-sharpening technique to yield a
60cm resolution color images (see Fig. 1) [11].
It has been analyzed in theory and shown by the numerical
example that DSmT does not work well for the change
detection, especially when the number of sources is larger
than 2. So only DER will be applied in this experiment. The
unsupervised clustering method ECM (Evidential C-Means),
detailed in [12], is applied for the image classification only
by using the radiometric point of view. ECM is adapted to the
classification of uncertain data, and the imprecise classes can
be acquired by ECM. The membership about the classification
of each pixel acquired by ECM are directly used as bbas here.

4 ={Gray area: road(w5 )or incomplete building(w6 )}

5 ={White area: white building(w7 )or ruin(w8 )}
={Ignorance}
The other tuning parameters are defined by: Maximum number of iterations T = 20, weighting exponent for cardinality
= 2, weighting exponent = 2, termination threshold
= 1. This classification technique has been used here
because it is unsupervised and the results can be directly
used as the mass functions (bbas). Nevertheless, any kind
of classifiers (including supervised ones) may be used at this
level. For the decision making, we take the criteria that true
hypothesis gets the maximum of pignistic probability.
The time analysis of the three images is mainly taken
for detecting the important change occurrences to evaluate
the damage, and we just pay attention to some particular
transitions rather than all the possible transitions. So the
constrained model of DER is applied here. The state
transitions considered from the image Fig. 1-(a) to Fig. 1-(b)
includes the change occurrences from the gray building to
ruin as t3,5 (ti,j = ti ,j ) and from incomplete building
to white building as t4,5 , and all invariable transitions as
t1,1 , t2,2 , t3,3 , t4,4 , t5,5 . The change occurrences from gray
building to ruin as t3,5 and from gray building to bared soil
as t3,2 which means the ruin has been cleaned, and all the
invariable transitions as t1,1 , t2,2 , t3,3 , t4,4 , t5,5 are involved
from the image Fig. 1-(b) to Fig. 1-(c). Therefore, the
constrained available transitions in the three images are given
by
t1,1,1 , t2,2,2 , t3,3,3 , t4,4,4 , t5,5,5 , t3,5,5 , t4,5,5 , t3,3,5 , t3,3,2 .
All the other possible transitions considered useless here are
defined as empty sets through constrained model in DER.
Remark: The transitions t3,5 and t4,5 may involves many
kinds of possible change occurrences, but some of the change
occurrences are unavailable according to our prior knowlM
edge. The constrained transitions can be defined by t3,5 =
M
tw4 ,w8 , t4,5 = tw6 ,w7 .

The combination results are shown as Fig. 2. Some change
occurrences are considered more important than the invariable
transitions for the evaluation of the disaster, and they are
extracted in Fig. 3.
The notations of transitions from 1st to 2nd image are:
t1,1 : {Green area}, t2,2 : {Blue area}, t3,3 : {Gray area},
t4,4 : {White area}, t5,5 : {Dark red area},
M

t3,5 = tw4 ,w8 : {Red area}, t4,5 = tw6 ,w7 : {Dark yellow area}.
t3,5 and t4,5 correspond to actual changes and they are linked
to damage mapping (see Fig. 3). Fig. 3-(a) focus on those 2

981

classes. The notations of transitions from 2nd to 3rd image:

illustrates the interest and the efficiency of DER for changes

detection and estimation. The DER fusion of images can well
detect the change occurrences and classify the invariable areas.
Our further research works will concern the extension of DER
for working with other DSm fusion rules, etc.

t1,1 :{Green area}, t2,2 : {Blue area}, t3,3 : {Gray area},

t4,4 :{White area}, t5,5 : {Red area}, t3,2 : {Cyan area},
M

t3,5 : = tw4 ,w8 : {Purple area}.

t3,5 , t3,2 corresponds to changes induced after the earthquake.
The notations of transitions through the 1st, 2nd and 3rd
images:
t1,1,1 : {Green area}, t2,2,2 : {Blue area},

Acknowledgements
This work is supported by China Natural Science Foundation (No.61075029) and PhD Thesis Innovation Fund from
Northwestern Polytechnical University (No.cx201015).
R EFERENCES

t3,3,3 : {Gray area}, t4,4,4 : {White area},

t5,5,5 : {Dark red area}, t3,3,2 : {Cyan area},
M

t3,3,5 = tw4 ,w4 ,w8 : {Purple area},

M

t3,5,5 = tw4 ,w8 ,w8 : {Red area},

M

t4,5,5 = tw6 ,w7 ,w7 : {Dark yellow area}.

We show more interest in the variable transitions than
the invariable transitions, since the variable transitions reflect
important change occurrences which is very valuable in the
evaluation of disaster. As we can see, some buildings were
entirely destroyed by the earthquake as represented by red
color mainly on the left side of the image and some incomplete
buildings has been completed represented by the yellow color
in Fig. 3-(a), which reflects the disaster mainly happened
on the left side. In Fig. 3-(b), the purple area indicates
that another normal building also collapsed, and the cyan
area means the run has been cleaned. Fig. 3-(c) shows the
sequential transitions of the three images, and it represents all
the particular change occurrences through the three images.
These fusion results can be helpful for the disaster evaluation
in the real application. There are some few false alarms of
change detections which are mainly due to the difference in
the geometry of the acquisitions. If some ancillary information
are available, these noisy changes can be reduced.
IV. C ONCLUSIONS
A Dynamic Evidential Reasoning (DER) approach has been
proposed for the change detection in the multi-temporal remote
sensing images. DER approach starts with the sequential
construction of the power set of admissible state transitions
taking into account, if necessary, some integrity constraints
representing some known unacceptable (impossible) transitions. Based on a particular rule-based algebra, the mass
of belief of the change occurrences can be computed using
two different combination rules the DERf or DERDS rule.
DERf rule in the free model works under the condition that
no prior knowledge about the change occurrences is known
with a great computation burden. DERDS rule in constrained
model is preferred when some constraints on the impossible
change occurrences is available to get better fusion results
with less computational complexity. Several simple numerical
examples were given to show how to use DER and to show
its difference with classical fusion approaches. Finally, an
experiment about the fusion of multi temporal satellite images

[1] A. Bouakache, A. Belhadj-Aissa, and G. Mercier, Satellite image fusion using Dezert-Smarandache theory, in
Advances and Applications of DSmT for Information
Fusion, F. Smarandache and J. Dezert, Eds. ARP, 2009,
vol. 3, ch. 22.
[2] G. Mercier, G. Moser, and S. Serpico, Conditional
Copula for Change Detection on Heterogeneous SAR
Data, IEEE Trans. Geosci. Remote Sensing, vol. 46,
no. 5, May 2008.
[3] G. Shafer, A Mathematical Theory of Evidence. Princeton Univ. Press, 1976.
[4] F. Smarandache and J. Dezert, Advances and Applications of DSmT for Information Fusion V1-3. Rehoboth,
USA: American Research Press, 2004-2009.
[5] S. Corgne, L. Hubert-Moy, and J. D. et al, Land cover
change prediction with a new theory of plausible and
a pradoxical reasoning, in Advances and Applications
of DSmT for Information Fusion, F. Smarandache and
J. Dezert, Eds. Am. Res. Press, Rehoboth, Jun. 2004.
[6] S. Hachicha and F. Chaabane, Application of DSM
theory for SAR image change detection, in Proceedings
of 2009 16th IEEE International Conference on Image
Processing (ICIP), Nov. 2009, pp. 37333736.
[7] W. Liu, Analyzing the degree of conflict among belief
functions, Artificial Intelligence, vol. 170, no. 11, pp.
909924, 2006.
[8] A. Martin, A. L. Jousselme, and C. Osswald, Conflict
measure for the discounting operation on belief functions, in Proceeding of Fusion, Germany, 2008.
[9] P. Smets, Decision making in the TBM: the necessity
of the pignistic transformation, Int. Jour. Approx. Reasoning, vol. 38, pp. 133147, 2005.
[10] , The combination of evidence in the transferable
belief model, IEEE Trans. Pattern Anal. Machine Intell.,
vol. 12, no. 5, pp. 447458, 1990.
[11] R. Andreoli, H. Yesou, and N. T. et al, Exploitation
en crise et post crise de donnees satellites haute et
tr`es haute resolution pour la cartographie de degats de
seismes. cas de Bam, Boumerd`es et Al Hoceima, in
Sirnat, Montpellier, Mar., 1011, 2005, (in french).
[12] M. H. Masson and T. Denoeux, ECM: An evidential
version of the fuzzy C-means algorithm, Pattern Recognition, vol. 41, pp. 13841397, 2008.

982

(a)

(a)

(b)

(b)

(c)

(c)

Figure 1. Multi-temporal QuickBird satellite images acquired before and

after an earthquake. (a) before image: 04/22/2002, (b) after image: 05/13/2003,
c
(c) latter acquisition: 06/13/2003. Dataset Boumerdes Copyright
SERTIT,
2009, distribution CNES.

Figure 2. Fusion results of classified images by DERDS . (a) Fusion of the

1st and 2nd image, (b) Fusion of the 2nd and 3rd image, (c) Fusion of the
1st, 2nd and 3rd image.

983

(a)

(b)

(c)
Figure 3. Significant damage map extracted from DER decision of Fig. 2).
(a) Changes from the 1st to 2nd classified image, (b) Changes from the 2nd
to 3rd classified image, (c) Changes through the 1st, 2nd and 3rd classified
images.
984